Hi @박 내석 ,
Thanks for sharing your questions.
I could not find a documented maximum relative error bound for the GetProcessTimes exit timestamp compared with a later GetSystemTimePreciseAsFileTime reading in the same Windows guest. Your successful wait establishes the execution ordering you described; the unresolved issue is the timestamp error model.
GetProcessTimes documents the returned value as the process's exit time, represented in FILETIME units. It does not specify:
- The exact clock-sampling or timestamp-storage point within process termination.
- The kernel clock-reading routine or interpolation algorithm used.
- The rounding or truncation rule.
- A maximum error relative to a subsequent precise UTC reading.
The 100 ns representation therefore does not establish 100 ns accuracy. PsGetProcessExitTime describes system time as typically updated approximately every ten milliseconds. That wording supplies neither a maximum exit-timestamp error nor a documented implementation relationship sufficient to transfer a bound to GetProcessTimes.
There is a stronger related statement in the KeQuerySystemTimePrecise documentation: KeQuerySystemTime is described as accurate within a system clock tick, while the precise routine uses the performance counter for interpolation and is described as accurate within a microsecond. However, those statements do not identify which routine supplies the process exit timestamp or when it is sampled. They therefore do not establish a "one tick plus one microsecond" bound for this comparison.
The submicrosecond precision documented for GetSystemTimePreciseAsFileTime also does not establish the uncertainty of the historical exit value. The timestamp difference includes the actual interval between the represented events as well as any differential timestamp error. System-time adjustments can affect that comparison even within one guest. The System Time documentation notes that system time can be adjusted forward or backward when Windows synchronizes with a time source and advises against comparing system time readings to determine elapsed time.
I found no guarantee in these sources specific to build 26100 or to the managed termination scenarios described. The missing patch identity and native termination trace limit implementation-specific conclusions, but do not change what the public documentation establishes. The ten observed inequalities remain valid observations without supplying a maximum tolerance.
A stronger guarantee would require a specification identifying the exit timestamp's sampling point, clock source, rounding behavior, and error bound, together with its applicable builds, termination paths, and clock-adjustment assumptions. Implementation analysis could establish behavior for a particular configuration but would need separate justification before being generalized.
And if you found my support helpful or informative, I'd really appreciate it if you could follow this guide so others with a similar issue can benefit as well.
Thank you.